This paper develops a task-based framework to measure occupational exposure to large language models (LLMs) in China and examines heterogeneity across occupations, workers, firms, and industries. Using the Chinese Occupational Classification Dictionary and a rubric-based evaluation of detailed task descriptions, we construct occupation-level exposure measures that incorporate task weights. We combine these measures with labor survey data and online vacancies from Chinese listed firms. The results show that exposure is highly uneven, with higher exposure concentrated in knowledge-intensive and information-processing occupations and among highly educated and high-income workers. We further introduce an exposure-dispersion measure to capture within-occupation task heterogeneity. Occupations with similar average exposure may differ substantially in their scope for task restructuring and reallocation. These findings suggest that LLMs reshape labor demand less through uniform occupational substitution than through uneven task-level exposure within occupations.